Information Technology Reference
In-Depth Information
Devices built on biological principles of information processing must have these
capabilities, which will be of great importance for the industry of information
processing.
Without attempting to outline in detail the history of neurocomputers, exten-
sively treated in [], let us focus only on the main factors characterizing this
important and rapidly developing field.
Today there are three active development tracks of neurocomputers:
First, emulators, i.e., systems based on digital von Neumann computers,
implementing typical neural network operations at the software level.
Second, neuroaccelerators, neural network systems implemented on the basis of
universal digital computers in the form of expansion cards. They may be both
“virtual” (compatible with expansion slots of standard PCs) and “external,”
connecting to the host computer via a specific interface or bus.
Finally, neurocomputers, employing specialized neural chips which execute all
operations in the neural network logical basis.
Without touching on neuroemulators and neuroaccelerators, we will mention
briefly the basic principles of construction and functional characteristics of
neurocomputers.
A special circuitry was developed in neuroinformatics for describing algorithms
and designing devices, in which elementary devices are combined into networks
designed to solve specific problems.
The following basic elementary devices are used (Fig.
4.4
):
Adaptive adder which computes the scalar product of the input vector
x
(i.e.,
information coming from all neurons) with the parameter vector
Nonlinear signal converter, which receives a scalar signal
x
and converts it into a
given function
f
(
x
)
Branch point, which is used to send the incoming signal to multiple addresses
a
1
x
output
signal
input
signal
a
2
∑
x
(
x
)
˕
x
x
x
˕
ʱ
n
x
a
n
∑
(
x
,
a
) =
x
i
i
i
=1
a
nonlinear
converter
branching
point
adaptive adder
branching
point
a
a
0
input
signal
∑
˕
ʱ
a
n
nonlinear
converter
formal neuron
Fig. 4.4 Circuitry of neurocomputing
